On September 10, local marketing platform Uberall published something the AI-search category rarely produces: an actual measurement. Across ChatGPT, Gemini, Claude, Grok and Perplexity, the company collected more than 120,000 AI business mentions spanning 3,793 physical locations in US cities from Chicago to New York, covering dentists, restaurants, grocery stores, hotels and banks. The finding that should stop you: in four of the five verticals, star ratings do not predict whether a business gets recommended by an AI model at all. Review volume does. For dentists, no star rating on any platform reached statistical significance.
The timing is not an accident. Over the past 18 months AEO/GEO went from a buzzword to a category with a price tag. Profound closed a $96 million Series C at a $1 billion valuation in February 2026, bringing four rounds to roughly $154.5 million. Scrunch AI raised a $15 million Series A and was then acquired by Sitecore for $225 million in June. Amsterdam-based Promptwatch took a €6 million seed. Almost all of that capital points at enterprise brand visibility inside AI answers. None of it answers what a twelve-table restaurant should do on Monday morning. Meanwhile the incumbent local players are repositioning: Uberall has raised around $174 million, covers close to 700,000 locations for more than 1,500 customers, and this study is the ammunition for its agentic product, UB-I.
Line the peers up and the split is obvious. Yext sells listings syndication, with published small-business tiers running roughly $199 to $999 per location per year. Birdeye sells review generation at roughly $299 to $449 per location per month, annual billing. Profound sells enterprise prompt monitoring and barely touches local at all. The category is moving out of its funding-expansion phase and into consolidation — Sitecore swallowing Scrunch was the first signal, local listings vendors folding GEO into existing subscriptions is the second.
For small operators, the value of this dataset is not that its conclusions are novel. It is that someone finally published thresholds: how many reviews, how many photos, how many attributes it takes to cross the line where a model says your name out loud. Below: the numbers, actions for three roles, a vendor comparison, and a route that costs no subscription at all.
What the study actually found
Uberall grouped the observable signals into four buckets it calls BARS — Business data, Authority, Review, Social. The important structural point is that some signals decide whether you get mentioned, and a different set decides how often.
- Business data. A filled-out Google Business Profile description tripled mention rates for grocery stores. Moving from 6–10 attributes to 31–50 took hotel mention probability from 22% to 94%. Photo count is the single strongest predictor of restaurant mention frequency — top-mentioned restaurants carry roughly three times more photos — and in dental it is the only signal that predicts both whether a practice is mentioned and how often.
- Authority. Brands with 30 or more news mentions saw a 15-fold frequency increase in banking and a 100% mention rate in grocery. Banks appearing on three or more editorial platforms saw a 13-fold increase, and Michelin recognition showed up in 94.5% of Perplexity's restaurant responses. Conversely, brand size — store count, practice count, deposit share, room supply — was a poor predictor across all five verticals. Independent restaurants and dental practices out-mention chains.
- Review. Restaurants past 1,000 Yelp reviews hit a 93.3% mention rate. Grocery stores past 500 Yelp reviews hit 100%; under 10 reviews, 13.3%. Dental practices past 1,000 Google reviews hit 92.9%, and mentioned practices averaged 643 reviews versus 253 for those never mentioned. Banks invert the intuition entirely: higher aggregate Yelp and TrustPilot ratings correlate negatively with mention frequency. Hotels are the lone exception, where GBP star rating beats review count.
- Social. Facebook follower count lifts the probability of being mentioned — mentioned dental practices had nearly five times the followers of unmentioned ones. Instagram amplifies frequency instead: restaurants with strong Instagram and Yelp presence were mentioned close to seven times more often, and for boutique hotels Instagram outranked even GBP and editorial signals.
Model personalities got quantified too. Gemini, cross-referencing live Google Maps data, surfaced eight times more unique restaurants than ChatGPT on the same queries. ChatGPT produced the shortest, stickiest shortlists — and the highest hallucination rates across verticals. Claude was the most conservative, largely refusing to name healthcare providers. Grok cited chef credentials and Instagram content more than any other model. Perplexity searched live, cited sources, and generated the most mentions per run.
What to do this week
Owners and operators
- ☐ Get one number first: how many reviews, photos and attributes your profile has right now. Under 100 reviews and 50 photos, GEO is not your problem yet.
- ☐ Shift part of your budget from raising the rating to raising review count — without loosening service standards. Rating still decides whether a human books.
- ☐ One local press or editorial-list placement beats three months of self-published content.
Marketing and SEO practitioners
- ☐ Build a monthly prompt panel: 20 fixed local-intent questions, run once against each model, logging whether you appear and in what position.
- ☐ Upload photos in steady increments rather than dumping 500 at once — the study is explicit that cadence signals an active business.
- ☐ For multi-location brands, fix the weakest 20% of locations before optimizing the top 20%. Threshold signals pay the most at the low end.
Developers and agencies
- ☐ Pull descriptions, attributes and photo counts for every location via the Google Business Profile API and score completeness.
- ☐ Fill in review counts via the Yelp Fusion API or the local equivalent, with threshold alerts at 1,000 (restaurants), 500 (grocery) and 1,000 (dental).
- ☐ Schedule the prompt panel: a Laravel scheduler plus queue hitting several model APIs, results into MySQL, one trend chart on the front end — that chart is a billable monthly report.
Vendor comparison
| Tool | Positioning | Local AI coverage | Public pricing | Best fit |
|---|---|---|---|---|
| Uberall | Multi-location listings plus UB-I agentic profile repair | Authors of the study; covers all four BARS factors | All three tiers require a sales call | Chains above 50 locations |
| Yext | Listings syndication across endpoints | Strong on data consistency, weaker on reviews and social | SMB tiers roughly $199–$999 per location/year | 1–10 locations focused on accuracy |
| Birdeye | Review generation and reputation | Attacks the review-volume factor directly | Roughly $299–$449 per location/month, annual | Single store to mid-size chain |
| Profound | Enterprise AI answer monitoring | Brand-level prompt tracking, not local fields | Enterprise quote | National brand marketing teams |
| Self-hosted (below) | APIs plus scheduler plus dashboard | Covers monitoring and completeness scoring | Model API costs only, often under $10/month | 1–20 location SMBs |
What the study does not say out loud
First: this is vendor-sponsored research whose conclusions happen to equal its product roadmap. The byline is labelled sponsored by Uberall, and the "three things to do this month" map line by line onto UB-I's feature list. More importantly, everything reported is correlation, not causation. Businesses with many reviews also tend to be older, better photographed and previously covered by media. Isolating "review volume" as a factor may simply be detecting a proxy for "this place was already popular."
Second: "chase volume, ignore rating" is the most dangerous sentence an SMB can act on. Models read counts; customers read stars. Pour everything into volume and conversion dies first. Worse, aggressive review acquisition runs straight into Google and Yelp policy on gating and incentivized reviews. Get flagged for manipulation and the whole batch gets wiped — that is not back to zero, that is negative.
Third: the photo advice contradicts the current risk environment. The study tells restaurants and hotels to push past 2,000 photos, but business-profile photo fields are exactly where abuse has concentrated lately. Padding the count with generated imagery may lift mentions short-term and then collect both platform enforcement and in-person expectation gaps.
The no-subscription route
A one-to-twenty location operator does not need a four-figure monthly contract. Three phases:
- Phase 1 (weeks 1–2), baseline. Export description length, attribute count, categories and photo count per location via the Google Business Profile API into a spreadsheet. Record Google and Yelp review counts by hand. That sheet is your baseline.
- Phase 2 (weeks 3–6), clear the thresholds. Prioritize by the study's own numbers: attributes above 30, description fully written, 10–20 photos uploaded weekly. Turn review requests into a fixed post-transaction step — no filtering, no incentives.
- Phase 3 (week 7 onward), measure. A scheduled script hits ChatGPT, Gemini and Perplexity APIs with 20 local-intent prompts on the first of each month, writes "mentioned yes/no" and rank position to a database, and renders a line chart. Combined API cost is typically less than a cup of coffee.
FAQ
Does a US study transfer to other markets?
The signal structure transfers; the platform weights do not. Where Yelp dominates in the US, other markets run on Google reviews, Facebook pages and local forums. The four BARS factors stay general: data completeness decides whether you are mentionable, authority and review volume decide how often.
Reviews first or photos first?
Completeness first. The study classes business data as an entry-ticket signal — an incomplete profile gives the amplifiers nothing to amplify. Going from 6–10 attributes to 31–50 moved hotel mention probability from 22% to 94%, the highest-leverage single cell in the entire dataset.
Do I need to optimize for all five models?
No. Pick the two your customers actually use. Gemini consumes live Google Maps data, which maximizes the return on profile maintenance. Perplexity retrieves live and cites sources, so it is the one you win by having genuinely good site content.
Do AI mentions turn into revenue?
There is no clean attribution path today. Most assistants send no referrer, so GA4 logs the visit as direct. The practical workaround is a dedicated landing page with its own UTM and phone number, and measuring calls rather than sessions.
Is this worth doing for a single location?
Yes, and it is cheapest there. The study found independent restaurants and dental practices out-mentioning chains, because threshold signals — photos, attributes, review counts — require consistency, not scale.
My take
My read runs against the category's own story: "local GEO" will not survive 18 months as a standalone SaaS line item. Translate BARS back into older language and you get profile completeness, external authority citations, review signals and social activity — which is the prominence-plus-relevance stack Google Local has shipped for a decade. The models look alike because they read the same layer: Google Business Profile, Yelp, editorial lists, social. When the underlying sources overlap this heavily, "optimizing for AI" and "finishing your local SEO" converge into one job, and the $300-a-month monitoring layer in between compresses into a tab inside someone else's dashboard. Sitecore paying $225 million for Scrunch is the start of that compression, not proof the category matured.
For a Laravel-plus-Flutter studio like ScriptWalker, the opportunity sits downstream of the monitoring tools, not inside them. Two productizable services: (1) a local AI visibility monthly report — profile completeness scoring via the Google Business Profile API, scheduled multi-model prompt runs, PDF out of a Laravel admin, sold as a retainer to 5–30 location chains; and (2) a review momentum engine — a Flutter staff app that fires a compliant review invitation at checkout, with a backend tracking monthly review growth per location against the thresholds above. Neither requires touching model training. Both deliver a data pipeline and an operating routine, which is exactly where an agency beats a subscription.
Sources
- Primary: Uberall — 5 AI Models, 4 Signals: What 120K Mentions Reveal About Multi-Location SEO (2026-09-10)
- Primary: Uberall blog — Visibility in ChatGPT and LLMs
- Primary: Yelp — Yelp and Hatch Advance Voice AI with OpenAI's GPT-Live-1 (2026-09-10)
- Primary: Profound — $20M to pioneer Answer Engine Optimization
- Primary: Google Business Profile API docs and Yelp Fusion API docs
- Third party: Fortune — Profound raises $96 million (2026-02-24)
- Third party: MarTech Cube — Scrunch AI raises $15M Series A
- Third party: PPC Land — Yelp Host past one million calls
- Third party: Crunchbase — Uberall funding and scale
- Third party (pricing roundups): Yext Pricing 2026, Birdeye Pricing 2026
- Further reading: Google DeepMind — Autoregressive Ranking (PDF), the longer-term variable in ranking architecture